Fig. 1

The overall framework of the AutoCoV model. a Preprocessing of SARS-CoV-2 sequences: We transform the virus sequences into a k-mer vector. After frequency normalization and information theory based k-mer filtering, we obtain an informative k-mer frequency matrix as inputs for AutoCoV. b The architecture of AutoCoV: It consists of three modules for learning the spatial and temporal patterns of SARS-CoV-2. Auto-Encoder Network generates latent representations that reconstruct the input matrix. Classifier Network guides the latent representations to identify the spatial and temporal patterns. Center Loss module complements the Classifier Network to create a more dense and well-separated embedding space. c The output of AutoCoV: The embedding space generated by AutoCoV aims to imply the spatial and temporal patterns of SARS-CoV-2